Upper extremity hemiparesis has a profound and lasting negative impact on quality of life and independence in activities of daily living for stoke survivors, yet, despite many investigations, there are no gold standard treatments in current clinical practice. We have carried out a progressive research program that has demonstrated the effectiveness of robotic therapy in chronic stroke patients, and then optimizing the type of robotic therapy and associated therapy that these patients receive. But gains in the chronic group are modest on average, and despite an intuition that earlier intervention will be beneficial, there is little data to confirm that. Therefore, we propose to build upon our expertise with robot-assisted training for chronic stroke impairment in order to evaluate the potential of predictive models that would select patients with better chances of making functional gains. Our clinical hypothesis is that kinematic and physiological biomarkers of recovery potential exist, and include baseline motor ability in both functional tasks and in the robotic environment, and measures of corticospinal tract effectiveness determined by transcranial magnetic stimulation. Functional and anatomical measures of connectivity, and plasticity relevant genotype are secondary biomarkers to test. Time after stroke is another promising marker to test, with strong implications for clinical practice. We also propose a secondary mechanistic hypothesis that maladaptive transhemispheric cortical inhibition will be altered by the intervention. Specific Objectives: Create a predictive model of function and disability following the intervention. Measure the effects of 12 weeks of robot-assisted therapy and transition to task training at > 6 months after stroke. Create a model that predicts clinically meaningful change in Fugl-Meyer in response to the intervention and test the validity of the model. Determine the effect our hybrid method of training has on interhemispheric inhibition by using transcranial magnetic stimulation to study silent period and recruitment curve. Patients with moderate to severe arm dysfunction (based on Fugl-Meyer scores of 7 to 45) of >6 months duration who are medically stable and do not have contractures or other impairments that would interfere with training will be enrolled. 96 subjects will be assigned to a single study arm with a multiple baseline approach to ensure stability of measures. Evaluations will be conducted by an examiner who has no knowledge of the predictive model. The first four weeks of training will consist of wrist robot training sessions, te second four weeks will consist of planar robot training sessions and the third four weeks will consist of alternation wrist and planar training sessions. Robot training sessions will be 45 minutes in duration followed by 15 minutes of training on functionally relevant tasks (translation to task training (TTT)). Clinical evaluations will include Stroke Impact Scale (Primary Outcome), Fugl-Meyer Upper Extremity Motor Performance Section Test, Wolf Motor Function Test, Action Research Arm Test, and activity monitor of home arm use. Kinematic analysis will be conducted pre and post training. A majority of patients will be consented for TMS and MRI. In those subjects intrahemispheric inhibition will be determined at each outcomes measurement visit. MRI will be performed only at the baseline period, and will be used to define the lesion type, white matter integrity and functional connectivity of the corticospinal tract. At the conclusion of these aims we will have a method for clinical decision-making regarding intensive arm therapy late after stroke, and a better mechanistic understanding of how it works. The method will be disseminated throughout the VA medical system and beyond.